A well-trained algorithm only needs a mass spectrum to deduce the molecular structure of a new designer drug.  

 

KVCV-Drugs

 

Don’t ask how it’s possible, but just make use of it, suggests Princeton researcher Michael Skinnider, whose algorithm secured the $15,000 NOMIS & Science Young Explorer Award 2023.  

Surprisingly, designer drugs, also known as novel psychoactive substances (NPS), arise from the conventional structure of drug legislation. Only molecules with structures explicitly outlined in the law are prohibited. Modify one subgroup, and you create something that probably has a similar effect on the human brain, but is entirely legal until the blacklist is updated—allowing the same trick to be repeated. 

A logical countermeasure, comparable to the proposed EU ban of PFAS compounds, involves banning entire substance families based on their carbon skeleton, irrespective of subgroups. Some countries, including Belgium and Germany, took this step years ago. But in The Netherlands, the proposed adaptation of current legislation still awaits consideration by parliament.  

At present, it only becomes apparent that the drug mafia has come up with something new when users end up in the hospital with unknown symptoms. Skinnider cites an example of a modified cannabinoid causing ‘zombie behavior’. Adequate medical help requires swift determination of the molecule’s structure. Mass spectrometry is an obvious choice, but only provides certainty when a reference spectrum is available. If not, an additional NMR analysis is required, which is more labor-intensive.    

ChatGPT 

Skinnider suspected that large language models, the neural networks that also form the basis of ChatGPT, could extract more information from a mass spectrum. Using the SMILES protocol (simplified molecular-input line-entry system), he digitized a database of 1753 known NPS structures. Although insufficient for training an AI, Skinnider cleverly took advantage of the fact that a SMILES translation can start at any chosen atom, yielding various descriptions of the same structure. 

The trained model could then devise new designer drugs, producing millions of molecules that seemed to meet the functional requirements, such as ease of synthesis and ability to penetrate the blood-brain barrier. Surprisingly, most structures appeared only once or twice in the output, while others were suggested tens of thousands of times.  

The reason for this remains a mystery, but Skinnider’s hunch that his AI’s logic aligns with that of the drug underworld doesn’t seem far off. Among 194 NPS that subsequently surfaced in society, 176 were found among the favorite suggestions. Since Skinnider published his work two years ago in Nature Machine Intelligence, his AI has correctly identified at least one unknown drug based solely on the total molecular mass. When the additional details from the MS spectra are taken into account, its power is even greater: about forty structures of newly discovered NPS have been elucidated in this way. 

 

The award-winning essay by Michael Skinnider, Hallucinating Hallucinogens, Science (2023)

Original paper: Michael Skinnider, et al., A deep generative model enables automated structure elucidation of novel psychoactive substances, Nature Machine Intelligence (2021) 

More information about the NOMIS & Science Young Explorer Award.  

 

Onderwerpen